Pre-segmented handwritten digit recognition using neural networks

Summary form only given. Results of current research suggest that multilayer neural networks with local `receptive fields´ and shared weights can be applied successfully to presegmented handwritten digit recognition. It was demonstrated that handwritten digit recognition without segmentation problem is actually quite simple; even traditional techniques such as the k nearest neighbor (KNN) classifier can provide good performance. Back-propagation, radial basis function (RBF) networks, and KNN classifiers all provide similar low error rates on a large presegmented handwritten digit database. The effectiveness of these classifiers `confidence´ was also evaluated. The back-propagation network uses less memory and provides faster classification but can provide `false positive´ classifications when the input is not a digit. The RBF network generates a more effective confidence judgement for rejecting ambiguous inputs when high accuracy is warranted. The KNN classifier requires a prohibitively large amount of memory and is much slower at classification, yet has surprisingly good performance